基于视觉短语的商品图像检索

发布时间：2018-06-22 12:17

本文选题：网上购物 + 视觉图像检索　；参考：《大连理工大学》2013年硕士论文

【摘要】：随着网上购物的快速发展,网络上出现百万张,甚至上亿张的不同类别的商业商品图片。传统的基于关键字的搜索引擎已经逐渐不能满足用户的需要。如何有效的利用视觉搜索手段来提供给用户精确度高、方便快捷的可视化搜索引擎目前仍然是一个重要的,富有挑战性的任务。在图像检索领域,词袋模型被广泛应用,结合倒排文件索引方法索引和检索数据库图像能够大大减少匹配中的候选图像数目,实现更有效的响应。基于词袋模型的搜索图像检索中忽略了点特征之间的空间关系以及逻辑关系,降低了一定的检索准确性。本文中,我们在词袋模型的基础上,结合点特征SIFT与区域特征MSER的不同特性,得到具有更好表达性的视觉词汇,以及在空间和逻辑上有紧密关系的图像的视觉短语。在文中,我们将在MSER区域内频繁出现的SIFT特征对即认为两个特征之间具有紧密关系,记为视觉短语。在完成整幅图像的全局特征匹配前提下,考虑到用户在购买商品时不仅会对商品的颜色,质地等因素有特别要求,同时也会对局部的设计感兴趣,针对这一点,我们提出同时提取图像的颜色、纹理等特征,并加入用户交互功能,以提供用户标注感兴趣区域的接口,进而实施系统对查询图像局部区域的第二阶段匹配,最后两阶段检索结果融合以提高检索精度。在实验评估部分,我们检验了利用本文提出的方法来实现的检索系统的实验结果。分别通过全局匹配,和局部匹配方法,对多次实验结果进行比较,验证了视觉短语和交互阶段局部匹配方法的有效性和必要性。
[Abstract]:With the rapid development of online shopping, there are millions, even hundreds of millions of pictures of different categories of commercial goods on the Internet. Traditional keyword-based search engines have been unable to meet the needs of users. How to effectively use visual search methods to provide users with high accuracy, convenient and fast visual search engine is still an important and challenging task. In the field of image retrieval, word bag model is widely used. Indexing and retrieving database images with inverted file indexing method can greatly reduce the number of candidate images in matching and achieve a more effective response. In the search image retrieval based on the word bag model, the spatial and logical relations between the points are ignored, and the retrieval accuracy is reduced. In this paper, on the basis of the lexical bag model, we combine the different characteristics of sift and MSER, and obtain visual words with better expression, and visual phrases of images closely related in space and logic. In this paper, the sift feature pairs, which occur frequently in the MSER region, are considered to be closely related to each other as visual phrases. On the premise of completing the global feature matching of the whole image, considering that the user will not only have special requirements for the color, texture and other factors of the product, but also be interested in the local design, in view of this, We propose to extract the color, texture and other features of the image at the same time, and add the user interaction function to provide the interface for the user to annotate the region of interest, and then implement the second stage matching of querying the local region of the image. In order to improve the retrieval accuracy, the final two stages of retrieval results fusion. In the part of experimental evaluation, we verify the experimental results of the retrieval system implemented by the method proposed in this paper. Through global matching and local matching, the experimental results are compared to verify the validity and necessity of visual phrases and interactive local matching.
【学位授予单位】：大连理工大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP391.41

【参考文献】